Skip to content

About

No description, website, or topics provided.

Resources

Stars

32 stars

Watchers

0 watching

Forks

Latest commit

 

History

8 Commits

Folders and files

Repository files navigation

Hybrid Forcing

Long-Horizon Streaming Video Generation via Hybrid Attention with Decoupled Distillation

Ruibin Li1,2, Tao Yang2, Fangzhou Ai2, Tianhe Wu3, Shilei Wen2, Bingyue Peng2, Lei Zhang1
1The Hong Kong Polytechnic University   2ByteDance   3City University of Hong Kong


Hybrid Forcing consistently achieves state-of-the-art performance, and thanks to the hybrid attention design the our model achieves real-time video generation at 29.5 FPS (832x480) on a single NVIDIA H100 GPU without quantization or model compression.


demo.mp4

🔥 News

  • 2026.4.6 : Work in progress, adapt hybrid video-audio joint streaming generation based on LTX2/2.3.
  • 2026.4.6 : The paper, project page, and code are released.

Quick Start

Installation

conda create -n hybrid_forcing python=3.10 -y
conda activate hybrid_forcing
pip install -r requirements.txt
pip install git+https://github.com/openai/CLIP.git
pip install flash-attn --no-build-isolation

# for hopper gpu, try FA3 and use CUDA 12.8 for best performance
git clone https://github.com/Dao-AILab/flash-attention.git
cd hopper
python setup.py install

Download Checkpoints

hf download Wan-AI/Wan2.1-T2V-1.3B  --local-dir wan_models/Wan2.1-T2V-1.3B
hf download Wan-AI/Wan2.1-T2V-14B  --local-dir wan_models/Wan2.1-T2V-14B
hf download xxx xxx --local-dir checkpoints

Inference

python inference.py \
  --config_path configs/hybrid_forcing.yaml \
  --output_folder output/hybrid_forcing \
  --checkpoint_path  xxx \
  --data_path prompts/demos.txt \

Training

Stage 1: Self-forcing training.

This stage is compatible with Self Forcing training, so you can migrate seamlessly by using the self-forcing code to train an initial stage1 checkpoint for 1000 step.

Stage 2: Hybrid-Forcing training.

Set your wandb configs before training.

And then train DMD models:

  export WANDB_API_KEY=xxx
  
  torchrun --nnodes=8 \
        --nproc_per_node=8 \
        --master_port=xxx \
        --nnodes=xxx \
        --node_rank=xxx \
        train.py \
        --config_path configs/hybrid_forcing.yaml \
        --logdir logs/hybrid_forcing

We recommend training another 1000 steps.

Acknowledgements

This codebase is built on top of the open-source implementation of CausVid, Self Forcing, Rolling Forcing, SLA, and the Wan2.1 repo, thanks for their excellect work.

References

If you find the method useful, please cite

@article{li2026long,
  title={Long-Horizon Streaming Video Generation via Hybrid Attention with Decoupled Distillation},
  author={Li, Ruibin and Yang, tao and Ai, Fangzhou and Wu, Tianhe and Wen, shilei and Peng, Bingyue and Zhang, Lei},
  journal={arXiv preprint arXiv:2604.10103},
  year={2026}
}

About

No description, website, or topics provided.

Resources

Stars

32 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages